Details
-
Improvement
-
Status: Resolved
-
Minor
-
Resolution: Fixed
-
0.10.0
-
None
Description
In the MusicProfiles example, when creating the user-user similarity graph, an edge is created between any 2 users that have listened to the same song (even if once). Depending on the input data, this might produce a projection graph with many more edges than the original user-song graph.
To make this computation more efficient, this issue proposes adding a user-defined parameter that filters out songs that a user has listened to only a few times. Essentially, it is a threshold for playcount, above which a user is considered to like a song.
For reference, with a threshold value of 30, the whole Last.fm dataset is analyzed on my laptop in a few minutes, while no threshold results in a runtime of several hours.
There are many solutions to this problem, but since this is just an example (not a library method), I think that keeping it simple is important.
Thanks to andralungu for spotting the inefficiency!
Attachments
Issue Links
- links to